The conversation around AI-assisted software development is becoming more nuanced.
| Stage | What happens | Human role | Risk |
|---|---|---|---|
| AI-assisted coding | AI helps write functions, tests, refactors, boilerplate | Human designs and understands | Low to moderate |
| Agentic coding | AI reads codebase, edits files, runs commands, fixes errors | Human directs and supervises | Moderate to high |
| Vibe coding | Human prompts by intention/feel and accepts what works | Human is experimenting, not deeply reviewing | Very high if shipped |
| AI-generated, human-reviewed code | AI writes significant code, but humans do white-box review, testing, verification, ownership | Human remains accountable | Manageable |
| AI-generated, AI-reviewed, human-rubber-stamped code | AI writes code, AI reviews code, AI explains why it is fine, human approves without deep understanding | Human appears accountable but may not be meaningfully in control | Extremely high |
Initially, we had AI-assisted coding. A human designed the solution. AI helped write functions, tests, refactors, and boilerplate. The risk was manageable because the human still understood the system.
Then came agentic coding. AI could read the codebase, edit files, run commands, fix errors, and iterate across multiple steps. This is powerful.
But the risk also increases because AI is no longer just completing a line of code. It is making changes across the system.
In parallel, we now have what people call vibe coding. The human prompts based on intention or feel. The code seems to work. The UI looks fine. The happy path passes. But the internals are not deeply understood.
That is fine for exploration. It is dangerous for shipping.
But I think we are now entering an even more interesting phase: AI-generated, AI-reviewed, human-rubber-stamped code.
AI writes the code.
AI reviews the code.
AI explains why the code is acceptable.
AI may even generate the tests.
Then a human, under delivery pressure, gives approval.
At that point, the human may not really be in the loop in a meaningful way. The human may only be providing institutional permission.
The problem is not AI-generated code.
The problem is unowned code.
In the AI era, architecture thinking matters more, not less.
Testing discipline matters more, not less.
Human accountability matters more, not less.
Code review should not become a ritual. It should become more serious.
White-box review matters more, not less.
Speed is useful, but ownership is non-negotiable.
If AI writes the code, AI reviews the code, and a human approves it without deeply understanding it, then we may not be keeping humans truly in the loop.
We may be slowly moving towards humans becoming rubber stamps in the loop.
That is where the real risk begins.